The paper introduces a Mixture-of-Agents (MoA) methodology to enhance the capabilities of large language models (LLMs). The authors propose a layered MoA architecture where each layer consists of multiple LLM agents. Each agent uses the outputs from the previous layer as auxiliary information to generate its response. This iterative refinement process continues until a more robust and comprehensive response is achieved. The MoA framework leverages the collective strengths of multiple LLMs, improving their reasoning and language generation capabilities. The authors demonstrate that LLMs exhibit a collaborativeness phenomenon, where models tend to generate better responses when presented with outputs from other models, even if those outputs are of lower quality. This collaborativeness is leveraged in the MoA framework to iteratively enhance the generation quality. The selection of LLMs for each MoA layer is guided by performance metrics and diversity considerations to ensure effective collaboration and improve overall response quality. Comprehensive evaluations on benchmarks such as AlpacaEval 2.0, MT-Bench, and FLASK show that the MoA framework achieves state-of-the-art performance, outperforming GPT-4 Omni by a significant margin. The MoA approach not only improves response quality but also demonstrates cost-effectiveness, achieving high performance with lower computational costs compared to GPT-4 Turbo. The paper also provides insights into the internal mechanisms of MoA, including the role of proposers and aggregators, and the impact of model diversity and the number of proposers. The results highlight the benefits of using diverse LLMs and the importance of a sufficient number of proposals in improving response quality. Additionally, the paper discusses the trade-offs between cost and performance, showing that MoA can achieve high LC win rates at lower costs compared to GPT-4 Turbo. Overall, the Mixture-of-Agents approach demonstrates the potential to enhance the effectiveness of LLM-driven chat assistants and improve the interpretability of models, making AI more accessible and aligned with human reasoning.The paper introduces a Mixture-of-Agents (MoA) methodology to enhance the capabilities of large language models (LLMs). The authors propose a layered MoA architecture where each layer consists of multiple LLM agents. Each agent uses the outputs from the previous layer as auxiliary information to generate its response. This iterative refinement process continues until a more robust and comprehensive response is achieved. The MoA framework leverages the collective strengths of multiple LLMs, improving their reasoning and language generation capabilities. The authors demonstrate that LLMs exhibit a collaborativeness phenomenon, where models tend to generate better responses when presented with outputs from other models, even if those outputs are of lower quality. This collaborativeness is leveraged in the MoA framework to iteratively enhance the generation quality. The selection of LLMs for each MoA layer is guided by performance metrics and diversity considerations to ensure effective collaboration and improve overall response quality. Comprehensive evaluations on benchmarks such as AlpacaEval 2.0, MT-Bench, and FLASK show that the MoA framework achieves state-of-the-art performance, outperforming GPT-4 Omni by a significant margin. The MoA approach not only improves response quality but also demonstrates cost-effectiveness, achieving high performance with lower computational costs compared to GPT-4 Turbo. The paper also provides insights into the internal mechanisms of MoA, including the role of proposers and aggregators, and the impact of model diversity and the number of proposers. The results highlight the benefits of using diverse LLMs and the importance of a sufficient number of proposals in improving response quality. Additionally, the paper discusses the trade-offs between cost and performance, showing that MoA can achieve high LC win rates at lower costs compared to GPT-4 Turbo. Overall, the Mixture-of-Agents approach demonstrates the potential to enhance the effectiveness of LLM-driven chat assistants and improve the interpretability of models, making AI more accessible and aligned with human reasoning.